Jack Lanchantin, Sainbayar Sukhbaatar, Gabriel Synnaeve, Yuxuan Sun, Kavya Srinet, Arthur Szlam
Meta AI
Install droidlet from here this directory (NeuralMemory/droidlet/).
Install requirements using: pip install -r requirements.txt
Verified using the following PyTorch and CUDA versions:
PyTorch version: 1.10.2
PyTorch CUDA version: 11.3
Generate data locally. Config file=world.active.all.all.txt, num_samples=100
./data/gen_data.sh local world.active.all.all.txt 100
This will generate train and validation files in the data/data/ subdir. These are PyTorch files that can be loaded using torch.load('val.pth'). All of the text data is encoded using the GPT-2 tokenizer. The dataloader and collator is found in data/db_encoder/.
You can change the config file to one of the following choices: world.active.all.all.txt, world.geometric_only.txt, world.property_only.txt, world.active.temporal_only.txt.
To change the query types, change the config file to one of the files from data/configs/. Modify the config file to change the query distribution.
If you change the config file or the number of samples, be sure to modify the --simple-data-path arg to math the newly generated data.
This code is CC-BY-NC 4.0 licensed, as found in the LICENSE file.